-
Notifications
You must be signed in to change notification settings - Fork 89
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
ReadMe update and QATv2 Refactoring #339
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks good,
using incoming qatv2 branch,the KeyError is fixed, but when I run |
Hi, thanks for getting in touch. Could you please share the ckpt file(2024.11.27-223950/qat_best.pth.tar) with us? Also, please indicate which script you are using: "evaluate_cifar100_mobilenet_v2_0.5.sh" or "evaluate_cifar100_mobilenet_v2_0.75.sh." |
I ran |
Please update your evaluation script to include the QAT policy used during training by specifying it as If the QAT policy is not specified in the evaluation script, the default policy is applied, where QAT begins at epoch 30. Since your checkpoint is from epoch 19, failing to declare the correct policy will result in a mismatch: the model will still contain batch normalization layers, but the checkpoint will lack the corresponding parameters. P.S.: Another thing that can be missed is the -8 parameter in the evaluation script. The -8 parameter is intended for evaluating a quantized checkpoint. Since your current checkpoint is from QAT, you must remove this parameter to properly evaluate the model. After completing the synthesizer steps, you can reintroduce the -8 parameter to evaluate the fully quantized checkpoint. |
resolves #340
1- quantile function replaced with torch.quantile()
2- Improved Distributed mode compatibility
3- Fixed the issue caused when loading best_ckpt with compile mode